The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the agent to download and execute an opaque setup script (./setup) from a third-party GitHub repository (https://github.com/garrytan/gstack.git). This pattern of 'download then execute' bypasses standard package manager safety checks and allows the remote repository to execute arbitrary commands on the host system during installation.
[EXTERNAL_DOWNLOADS]: The installation process requires cloning a large (approx. 98 MB) repository and downloading additional binaries, including Chromium, which increases the attack surface and dependency on external infrastructure.
[INDIRECT_PROMPT_INJECTION]: The skill operates by performing a 'visual audit' of project UI code and automatically generating fixes. This creates a vulnerability surface where malicious content embedded in the source code being analyzed (e.g., in HTML comments or CSS) could influence the agent to perform unauthorized file modifications or Git operations.
Ingestion points: Project source code, UI components, and design files.
Boundary markers: None specified in the provided instructions to delimit untrusted project content from agent instructions.
Capability inventory: File system write access, Git commit/stash capabilities, and execution of local tooling binaries.
Sanitization: No evidence of sanitization or validation of the content discovered during the visual audit before it is used to generate code fixes.
[COMMAND_EXECUTION]: The skill uses shell commands to check for its own installation and to perform its primary function of modifying files and managing Git commits.

design-review-garrytan